ocaml-solo5/lib/hvf/ at main · gazagnaire.org/ocaml-solo5 · Tangled

gazagnaire.org / ocaml-solo5

0

Solo5 hvt ABI and MFT1/ABI1 note format as wire codecs

0

ocaml-solo5 / lib / hvf

4 files

at main

Thomas Gazagnaire solo5: update to Solo5 v0.12.0 and add a power-off watchdog 2w ago

solo5: Move slirp out of the tender Run slirp as a separate gateway process over a datagram socketpair so the tender can keep the strict no-network sandbox, with backend-specific poll hooks for the Solo5 POLL hypercall.

1 month ago

solo5: update to Solo5 v0.12.0 and add a power-off watchdog Solo5 v0.11 extended struct hvt_boot_info with ring-I/O feature negotiation, so the refreshed reference headers grow the boot-info codec by host_features, guest_features and net_ring. The tenders offer no features and a v0.11+ guest answers by falling back to hypercall I/O; such guests also reserve 2 MiB of BSS for ring write buffers, so the boot harnesses give the real unikernels more RAM. The tender watchdog returns, this time counting only time spent inside the guest: the budget covers a single vcpu-run entry with no VM exit, so a guest blocked host-side in a long POLL is never killed, the accounting flaw that got the old per-hypercall watchdog removed. Both backends cancel the run from a watchdog thread (hv_vcpus_exit on hvf, SIGURG on kvm, its syscalls allowlisted only when armed), and the boot harnesses assert the watchdog's own diagnostic for runaway guests instead of relying on shell timeouts.

2 weeks ago

ocaml-solo5: Add shared tender and KVM backend Move the backend-independent hvt tender into solo5-tender and add the Linux KVM backend alongside HVF. Replace the old generated C ABI bindings with Wire codecs and direct hot-path accessors, add a single production benchmark suite with probe coverage, and update the tests and package metadata for the new layout.

1 month ago

solo5: update to Solo5 v0.12.0 and add a power-off watchdog Solo5 v0.11 extended struct hvt_boot_info with ring-I/O feature negotiation, so the refreshed reference headers grow the boot-info codec by host_features, guest_features and net_ring. The tenders offer no features and a v0.11+ guest answers by falling back to hypercall I/O; such guests also reserve 2 MiB of BSS for ring write buffers, so the boot harnesses give the real unikernels more RAM. The tender watchdog returns, this time counting only time spent inside the guest: the budget covers a single vcpu-run entry with no VM exit, so a guest blocked host-side in a long POLL is never killed, the accounting flaw that got the old per-hypercall watchdog removed. Both backends cancel the run from a watchdog thread (hv_vcpus_exit on hvf, SIGURG on kvm, its syscalls allowlisted only when armed), and the boot harnesses assert the watchdog's own diagnostic for runaway guests instead of relying on shell timeouts.

2 weeks ago